THE ATTENTION AGENT / LOCAL GEMMA

Give your attention
a plan.

Source-linked laboratory notes. Visible uncertainty.
A clear handoff when you return.

Recorded Gemma examples · live browser decisions and timers

RESEARCH PREVIEW · NO EQUIPMENT CONTROL
START WITH A · INSPECT B · EXPLORE C

A source-supported wait. Inspect the evidence and manually confirm a ten-second demonstration.

YOUR PROTOCOL STAYS YOURS.LOCAL GEMMA · SOURCE-LINKED SUGGESTIONS · HUMAN-CONTROLLED TIMERSNO EQUIPMENT CONTROL
YOUR ATTENTION DESK

Your attention desk.

NO ACTIVE EXPERIMENT
The bench notes
ORIGINAL WORDS / NO GENERATED INSTRUCTIONS
SOURCE

Focused source-pair analysis preserves headings and original IDs. Unresolved targets and model abstentions require review. See the length-stratified evaluation; inspect every suggestion.

Up to 48 nonempty lines. Only synthetic or approved non-sensitive notes. Input is kept in browser/server memory.

Load the example to explore the workflow without a model.
The Attention Map
SOURCE ORDER · NOT A PREDICTED START SCHEDULE
LOADING RECORDED EXAMPLE
ActionFixed waitReview

Good science.
A little breathing room.

Every source line stays in view. No invented durations. No silent missing steps.

Ask BenchSloth
03 / INTELLIGENCE · EVIDENCE FIRST
Recorded Gemma output

Ask what needs attention, which note belongs where, or what the source leaves unanswered.

Current browser facts. Deterministic, no new model inference.

Gemma selects evidence; the source stays verbatim. Answers cannot start timers or change your plan. A valid citation does not prove a correct interpretation.

THE MODEL SUGGESTS CONNECTIONS. THE SOURCE DECIDES THE TIMER.Reload clears browser state. Use approved equipment timers.
READUNDERSTANDVERIFYPLANRESTRETURN
02 / THE METHOD

Keep the evidence.
Make room to think.

01 / SEE THE SOURCE

Keep the original words.

Every nonempty source line has an occurrence ID. Nothing is silently omitted, merged, or rewritten into an experimental instruction.

02 / CONNECT THE CONTEXT

Know what needs attention.

Ask Gemma to find the original evidence behind a question. Inspect its selection, uncertainty and real request before accepting an interpretation.

03 / CONFIRM THE WAIT

Your judgment stays central.

Only exact, rule-supported intervals get a timer option. Restrictions and ambiguous timing remain visible for human review.

04 / TAKE A BREATH

Return to the source.

A calm countdown keeps the supporting passage and next source line in view. It never means required monitoring can stop.

Trust is a method.
Not a decorative badge.

This is an experimental research organizer, not a validated protocol or laboratory controller. Prepared results are labeled; real inference uses local Ollama. BenchSloth has not yet demonstrated human time savings. Inspect the frozen evaluations and limitations in the project README before relying on a classification.

Does BenchSloth Actually Work?

NEW / 24 frozen semantic questions: combined candidate15/24 exact evidence answers versus14/24 rules, but13/24 exact QA workflows versus14/24 and8 unsupported evidence selections versus7. Retrieval's top3 contains all required evidence for16/16 answerable questions. No reliable model advantage is established. Model relationships remain experimental proposals requiring review.

Separate unchanged v0.5 planner:20 cases,22 reference links,1 extra unresolved association,16/18 eligible waits,0 observed unsupported offers. Strong scoped rules tie. Independent human n=0; no measured time savings.

Explore every case, answer and failure ↗ · Optional core source ZIP ↗

Human trials completed: 0. Research preview; source interpretation still needs human review.